Following the concepts presented on my post named Should you use FastAI?, I’d like to show here how to train a Multi-Task deep learning model using the hybrid Pytorch-FastAI approach. The basic idea from the Pytorch-FastAI approach is to define a dataset and a model using Pytorch code and then use FastAI to fit your model. This approach gives you the flexibility to build complicated datasets and models but still be able to use high level FastAI functionality.

Multi-Task Learning (MTL) model is a model that is able to do more than one task. It is as simple as that. In general, as soon as you find yourself optimizing more than one loss function, you are effectively doing MTL.

In this demonstration I’ll use the UTKFace dataset. This dataset consists of more than 30k images with labels for age, gender and ethnicity.

If you want skip all the talking and jump to the code, here is the [

1. Why Multi-Task Learning?

When you look to someone’s picture and try to predict age, gender and ethnicity, you’re not using completely different parts of your brain right? What I’m trying to say is that you don’t try to understand the image in 3 different ways, each one specific to each task. What you’re doing is using a single understanding that your brain makes of the image and then trying to decode that understanding into age, gender and ethnicity. And besides that, there is knowledge from gender estimation that might help on age estimation, there is knowledge from ethnicity estimation that might help on gender estimation and so forth.

So, why MTL? We’re confident that training a single model to do all tasks we’re interested in yields better results than training a model for each task. Rich Caruana summarizes the goal of MTL pretty nicely: “MTL improves generalization by leveraging the domain-specific information contained in the training signals of related tasks”.

As this is a practical tutorial for training MTL models I’ll not dig deeper into theory and intuition but if you want to read more, check out this amazing post from Sebastian Ruder.

#pytorch #artificial-intelligence #deep-learning #data-science #fastai #data analysis